LTX Director

If you've ever stared at a ComfyUI LTX graph and thought "this is a video tool, why am I editing JSON arrays to time a shot," LTX Director is the answer. It's the WhatDreamsCost pack's flagship: a full timeline editor that lives inside a single node, where your text, image, and audio segments are draggable blocks you trim and cut on a ruler instead of numbers in a widget. It's built for LTX 2.3, which is a first-frame/last-frame model with time-windowed conditioning - stock ComfyUI has no good way to say "this, then that, at this moment." Director is the community's solution, and with the 2.0 update (June 2026) it stopped being a keyframe helper and became an actual AI video editor.

The author's own framing is worth respecting: he spent a month and a half on 2.0 and it shows - IC-LoRA support, audio inpainting, Retake Mode, timeline save/load, all of it free. Community reception was strong (the 2.0 launch thread pulled hundreds of upvotes), and it's now a fixture of the LTX 2.3 tooling discussion alongside LTX Desktop.

How it works

Director is, at its core, Kijai's Prompt Relay idea wrapped in a WYSIWYG interface. Prompt Relay lets each video segment carry its own local prompt that only conditions its time window (via Gaussian penalty masks on cross-attention). Director gives you those segments as blocks on a timeline, then serializes everything into JSON (timeline_data) plus the local_prompts, segment_lengths, and guide_strength strings - all auto-managed, don't hand-edit them. At execution it parses that state, VAE-encodes your image/video segments into guide frames, builds a Prompt Relay conditioning with soft or hard segment boundaries, and generates an empty LTXV latent (or uses your optional_latent). With an audio_vae connected it produces an audio latent too, which is how LTX-2's synchronized audio gets into the mix.

One thing beginners trip on: duration_frames only sets the visual scale of the timeline - the actual frame count comes from the latent you connect. Change the latent length and the timeline rescales; change the widget and nothing about your generation changes.

The inputs that matter

• model and clip - your LTX model and its CLIP. These get patched, so you'll see them on the output side too.
• global_prompt - conditions the entire video and anchors persistent characters/scene context across segments. Learn this one; it's how you stop a character from mutating between shots.
• epsilon - Prompt Relay's penalty decay. Below ~0.1 you get sharp segment boundaries (the default 0.001); raise it to 0.5+ for softer transitions. This is the hidden lever for "the cuts feel too harsh."
• use_custom_audio / inpaint_audio - timeline audio vs. generated-from-scratch, and whether empty gaps get audio-inpainted.
• custom_width / custom_height / divisible_by - how guide images get resized to LTX-friendly dimensions (32-divisible).
• override_audio - pull audio from the IC-LoRA video instead of the track.

Outputs

The important ones: model (patched), positive (Prompt Relay conditioning), video_latent, audio_latent, guide_data and motion_guide_data (the keyframe/motion payloads), frame_rate, and combined_audio. The standard wiring is Director → LTX Director Guide (which consumes guide_data, motion_guide_data, the latent, and the model) → LTX Director Crop Guides → sampler. Simple text-to-video runs can skip the Guide, but any image or video segments route through it.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/WhatDreamsCost/WhatDreamsCost-ComfyUI

Or search WhatDreamsCost in ComfyUI Manager. The README is explicit about the catch: update ComfyUI-LTXVideo and ComfyUI-KJNodes to their latest versions first - "you cannot use this node without updating ComfyUI-LTXVideo!" If the manager list doesn't show v2.0.0 yet, install the nightly to get the latest. Download the example workflows from the repo; they're the fastest way to see a correct graph.

Gotchas

Retake Mode is officially beta and the author says it's "not potent enough" - he's mid-overhaul, so don't judge the node by it. The 2.0.4 update added model caching and dropped Prompt Relay overhead from plain t2v/i2v runs, so don't be surprised when things suddenly feel faster; the quoted 1.5–3x speedup is hardware-dependent, so if you land at the low end, that's normal. And keep in mind the honest community critique that applies to any big overhaul node: the convenience hides the parameters you need when output goes wrong - Director hides Prompt Relay's guts behind the timeline, so when a segment misbehaves, reach for epsilon and global_prompt before assuming the tool is broken.